Do workers discriminate against their out-group employers?

Evidence from an online labor market

Sher Afghan Asad

Iowa State University

Ritwik Banerjee

IIM, Bangalore

Joydeep Bhattacharya

Iowa State University

November 4, 2019

Abstract

Discrimination in labor markets is often thought of as being one-sided, driven mostly by employers toward theirout-group employees. This paper restricts attention to racial identity and studies possible discrimination in the “re-verse” direction, i.e., we ask, do workers discriminate (say, by under providing effort) for an out-race employerrelative to an otherwise-identical, own-race one? We postulate that the main outlet for such discrimination is viasocial preferences of workers towards their employers. To test our hypothesis, we run a well-powered model-basedexperiment on Amazon’s Mechanical Turk (M-Turk) using a piece rate design. We find that workers do not dis-criminate against black employers in this online environment in fact there is some evidence for pro-social responsetowards the black employers. The results indicate that, in the environment that we study, worker discrimination maynot be responsible for the low representation of blacks in leadership positions and self-employment opportunities.The results also suggest that worker discrimination may not be responsible for explaining employer discriminationat-least in our setting.

The authors are grateful for funding from the Russell Sage Foundation and from the Economics Department at Iowa State University. SherAfghan Asad is a Ph.D. candidate in the department of economics at Iowa State University. Ritwik Banerjee is an associate professor of economicsat Indian Institute of Management Bangalore. Joydeep Bhattacharya is a professor of economics at Iowa State University. The authors are thankfulfor the thoughtful comments and suggestions by Otavio Bartalotti, Michael Best, Stefano DellaVigna, Catherine Eckel, Elizabeth Hoffman, AlexImas, Peter Orazem, and Devin Pope.

1

1 Introduction

A large body of literature in economics has demonstrated that discrimination - whether it be racial, religious, ethnic or

gender in origin - is widespread in labor markets.1 In attempting to understand the sources of discrimination, almost

the entire literature investigates the issue on the premise that it is one-sided, driven largely by employers toward their

out-group employees.2 In this paper, we focus on racial identity and study possible discrimination in the reverse

direction, i.e., we ask, do workers discriminate on the intensive margin (say, by shirking or under-providing effort)

against an out-race employer relative to an otherwise-identical, own-race one?

Discrimination in labor markets arise in one of two main ways. Becker (1957) introduced the notion of taste-

based discrimination postulating that discrimination exists because of a prejudice/animus towards the members of

the disadvantaged group. Phelps (1972) theorized, instead, that discrimination might be statistical – an employer,

lacking information about a job-seeker’s productivity, forms beliefs about it based on the person’s group identity and

the aggregate productivity distribution of the group to which the person belongs.3 Economists theorize that it is these

tastes and beliefs which lead employers to discriminate against equally productive workers from disadvantaged group

in favor of workers from the dominant group. Applying the same explanations to the worker side it can be argued

that workers may discriminate against an employer from a particular group simply because they dislike (taste) them.

Similarly, they may discriminate because they harbor negative beliefs (statistical) about the way an employer from a

particular group may treat them (Craig & Fryer, 2018). As will become clear below, our experimental design preclude

the possibility of discrimination arising due to beliefs or statistical reasons. Any discrimination we intend to find is

taste-based that may arise via social preferences (or lack thereof) of workers towards their employer’s group identity.4

Just like employer discrimination, worker discrimination can manifest itself either on the extensive margin (e.g.,

low rates of application to a job from a certain group employer) or along the intensive margin (e.g., deliberate under-

performance when working for an employer from a particular group). Discrimination along either margin is emphati-

1See Riach and Rich (2002), Rich (2014), Bertrand and Duflo (2016), and Neumark (2018) for a survey of evidence on discrimination usingexperiments.

2Existing anti-discrimination legislation views the matter in the same one-sided way as well. ‘It shall be an unlawful employment practice foran employer . . . to discriminate against any individual with respect to his compensation, terms, conditions, or privileges of employment, because ofsuch individual’s race, color, religion, sex, or national origin.’ (Title VII, Civil Rights Act of 1964). Donohue (2007) survey on anti-discriminationlaw is entirely silent on the possibility of worker-to-boss prejudice.

3See Bohren, Haggag, Imas, and Pope (2019) for discussion on distinction between correct and incorrect beliefs.4Dominant views of prejudice from the psychology literature postulate that a) prejudice is an evolutionary phenomenon which makes group

membership an important component of one’s social identity (Neuberg & Cottrell, 2006; Tajfel, 1970) and b) subconscious, knee-jerk negativeassociations can be triggered by exposure to members of an out-group (Bertrand, Chugh, & Mullainathan, 2005; Cuddy, Fiske, & Glick, 2007).Within this paradigm, a worker may exhibit positive feelings of altruism and reciprocity and work harder for employers belonging to his group andnot for others.

2

cally costly and demoralizing to those being discriminated against: it may directly reduce the performance or profitabil-

ity of businesses run by disadvantaged group employers therefore possibly causing inefficiently-low entrepreneurship

by minority entrepreneurs (Levine & Rubinstein, 2017); it may lead to lower representation of minorities in the lead-

ership roles;5 it may lead to sub-optimal growth in the economy (Hsieh, Hurst, Jones, & Klenow, 2019). In this paper,

we explore whether workers actually under-perform for the disadvantaged group employers in the online environment,

and if they do that might provide one potential reason for the lack of success of blacks in the leadership positions or

the self-employment opportunities.

Another motivation for our question stems from the recent evidence in economic literature that shows that workers

under-perform when working for the opposite group employer. For example, Glover, Pallais, and Pariente (2017) find

that minority workers, when they work for implicitly biased managers, under-perform as compared to same workers

when they work for un-biased managers. However, we test whether workers may under-provide effort even in the

absence of bias from the employer or anticipation of bias from the employer. Similarly Benson, Board, and Meyer-ter

Vehn (2019) find that “productivity of white and Hispanic workers has greater mean and lower variance if hired by

same-race managers”, indicating that workers’ performance is influenced by the social identity of their hiring manager.

While Benson et al. (2019) interpret their findings as resulting from better screening of the applicants by the same race

manager, we explore how workers from one group may perform in the absence of role of selection into the job.

This independent preference-based discrimination by workers may help further explain the self-fulfilling prophecy

of discrimination in which a rational, unbiased employer expecting potential employees from a particular group to

exhibit bias (perform poorly), decide to discriminate against such people in favor of his own-group employees. Taking

this line of thinking further, preference-based discrimination by worker may imply that labor-market discrimination,

mostly understood by economists to be taste-based (Becker, 1957), may be somewhat statistical (Phelps, 1972) because

it is driven, not by animus of the employer, but by beliefs a rational employer holds about the animus of employees

from a particular group.

There is a recent limited literature which investigates the discrimination by subordinates (Abel, 2019; Ayalew,

Manian, & Sheth, 2018; Chakrabortyy & Serra, 2019; Grossman, Eckel, Komai, & Zhan, 2019). This literature

focuses on gender as the social identity and mostly finds belief based discrimination against the female leaders. Our

5According to Korn Ferry (2019), since 1955, there have only been 15 Black CEOs at the helm of Fortune 500 companies, today, there are onlyfour. Chakrabortyy and Serra (2019) finds that minority group leaders may select out of leadership roles for the fear of backlash by the workers.

3

is the first study that directly investigates the possibility of preference based discrimination by subordinates in the

context of race in United States.

Based on the discussion above, we design an experimental environment characterized by the following features:

i) contracts are incomplete, allowing workers to work as little or as much possible; ii) employer’s social identity can

be made salient to the worker; and iii) the only source of discrimination is via preferences, there is no room for belief

based discrimination; iv) the task has real cost of effort (unlike monetized cost in studies such as Charness, Rigotti,

and Rustichini (2007); Fershtman and Gneezy (2001)); v) the interaction between worker and employer is one-shot so

that confounding reputation effects (of the kind that naturally emerge in Glover et al. (2017) and Benson et al. (2019))

do not enter.

Inspired by Dellavigna and Pope (2018), we outline an experiment using subjects from Amazon’s Mechanical

Turk (M-Turk) and black and white student subjects from a major public university in the United States.6 The student

subjects are recruited to be “Employers” while M-Turk subjects are assigned as “Workers.” Workers do not choose

which employer to work for instead they are randomly assigned to an employer and then work on a real-effort task

which requires a worker to alternately press the ‘a’ and ‘b’ buttons on the keyboard for 10-minutes.7 The worker’s

performance is measured by the number of times the buttons are alternately pressed and the worker is informed

(truthfully) of the return to employer from the worker’s performance. For example, the employer earns 1 cent for

every 100 points scored by the worker where a point is scored after one successful ’a’-’b’ press. The contract is not

complete as it does not stipulate the exact number of points that the worker must score in the task. Reciprocity is

triggered by offering small monetary gifts to worker on behalf of the employer. Possibility of discrimination in social

preferences is allowed by revealing the race of the employer to the matched worker.

Our pre-registered experiment design is based on a simple structural model inspired by DellaVigna, List, Mal-

mendier, and Rao (2016) in which workers have social preferences toward their employer and maximize utility from

the provision of costly effort. The ten treatments we design help us identify the cost-function and social-preference

parameters (altruism and reciprocity) of our structural model separately for neutral, black, and white employers. The

treatments are presented in table 1. The piece rate treatments are aimed at eliciting the elasticity of effort with respect

6M-Turk is an online platform that allows researchers to recruit human workers for small tasks and has become very popular for experimentalresearch in economics (e.g. Kuziemko, Norton, Saez, and Stantcheva (2015) and Freeman, Halevy, and Kneeland (2019)).

7The employer do not get to make any strategic choices (such as wage offers, minutes of work, etc.) thereby eliminating most channels forstatistical discrimination by workers.

4

to return to effort, while altruism (reciprocity) treatments with neutral, black, or white employers allow us to estimate

altruistic (reciprocal) preference parameters of workers towards neutral, black, or white employers respectively.

We take the approach of revealing race indirectly via the revelation of skin color. To that end, “employer-subjects”

are videotaped while they read off a script explaining and demonstrating the “a-b” task for the workers. The camera

placement only captures the hand of the employer along with the movement of the fingers alternating ‘a’ and ‘b’ button

presses. Other identifiers, such as the face, are not shown in the video. Workers are required to watch the video by

their matched employer before they can continue with the task. .

Overall, our results indicate that workers do not exhibit disparate social preferences towards the employer race in

this online environment. If anything, workers are slightly pro-social towards black employers as compared to white

employers which is interesting given that we only recruit white workers for this study. The small pro-social response

towards black employers is partially driven by the workers with low implicit bias against blacks (as proxied by the IAT

score of the worker’s region).

Our “null result” goes in contrast to findings in discrimination literature in economics in which 97 percent of the

discrimination papers published in top economics journal find the significant evidence for discrimination (Bohren et al.,

2019). Ours is a well-powered, pre-registered experiment which could have detected preference-based discrimination

if it existed in this platform, the fact that we don’t find significant evidence for such discrimination, at-least in this

setting, is encouraging for the American society.

Our results are especially encouraging in the context of transition to gig-economy. With expanding gig-economy

(Katz & Krueger, 2019), our results seem to suggest that discriminatory preferences may be less likely to elicit them-

selves in the economic interactions in these settings. This may be essentially due to the nature of jobs on these

platforms which is quite distinct from “traditional jobs” in a sense that worker-employer engagement is short term and

there is very little possibility of associative distaste to emerge. In fact, the distaste parameter must be quite large for

the disparate outcomes to emerge in such a setting. In that sense, our results can be interpreted as the lower bound on

the worker’s distaste towards the black employers.

Our research relates to the broad literature on discrimination in the labor markets. We specifically look at worker’s

vertical discrimination towards their employers, complementing a larger literature about horizontal discrimination

between co-workers (e.g. Becker (1957) and Sasaki (1999)). This study is specifically related to experimental literature

5

on discrimination. See Neumark (2018) and Bertrand and Duflo (2016) for reviews of experimental research on

discrimination. This research also relates to literature on identity and social preferences at work (Akerlof & Kranton,

2005; Benjamin, Choi, & Strickland, 2010; Chen & Li, 2009). The main contribution of this study in this context is

the novel design that can estimate the social preference towards the employer’s group in a unique environment of an

online labor market.

The rest of the paper proceeds as follows. In Section 2 we present the model of behavior and come up with the

treatments to identify the parameters of interest. In Section 3 we present the experiment design. Section 4 summarizes

the data for this study. In Section 5 we present the results followed by structural estimation in section 6 and conclusion

in section 7.

2 Model and Treatments

We design our experiment by closely following the model in DellaVigna et al. (2016) and modifying it to model

discrimination from the workers’ side. In the setup workers take part in a real-effort task where they choose how much

effort to provide. In the model, a risk neutral worker works for an employer j, j ∈ {Neutral,Black,White}, and solves

the following problem:

max︸︷︷︸e j≥0

U j = max︸︷︷︸e j≥0

(F +(s+ρ j1Gi f t +α jv+ p)e j− c(e j)

)(1)

where e j is the number of points (on a button-pressing task) scored by worker when working for an employer j,

F is the fixed money paid for participating in the experiment, s captures the sense of duty, norm, intrinsic motivation,

and competitiveness of worker towards the task which is independent of the employer. ρ j is the reciprocity parameter

per unit of effort which kicks in whenever an employer j awards a gift to the worker, Gneezy and List (2006). 1Gi f t

is an indicator function which assumes a value 1 when a gift is rewarded by the employer, 0 otherwise. α j captures

the altruistic preference of worker towards employer j per unit of effort Becker (1974), given that v is the value of the

unit of effort to the employer. p is the piece rate per unit of effort. c(e j) is the cost of effort, which is assumed to be

the same for all workers.8 We assume the regularity conditions c′() > 0, c′′() > 0, and lime→∞c′(e) = ∞. Following

8We relax this assumption later and allow for the possibility of heterogeneous cost.

6

Dellavigna and Pope (2018) and DellaVigna et al. (2016), we analyze the optimality conditions assuming two different

functional forms for the cost of effort function : a power function and an exponential function i.e.

c(e) =ke1+γ

1+ γ(2)

and

c(e) =kexpγe

γ(3)

Power cost function (2) characterizes a constant elasticity of effort with respect to return to effort given by 1/γ ,

while exponential function (3) represents decreasing elasticity of effort with respect to return to effort given by

1/log(r/k), where r is the return to the effort. Workers’ effort at different piece rates can be used to identify and

structurally estimate both the parameters of the cost of effort functions, namely, k and γ .

Solving 1 leads to the following solution (when interior):

e?i j = c′−1 (si +ρi j1Gi f t +αi jv j + p)

(4)

An assumption of the power cost function yields the following optimality condition:

e?i j =

(si +ρi j1Gi f t +αi jv j + p

k

)1/γ

The exponential cost function on the other hand leads to the following first order condition:

e?i j =1γ

log(

si +ρi j1Gi f t +αi jv j + pk

)

We make a simplifying assumption that the workers of type i are homogeneous given a treatment i.e. they will

make the same effort choice in a given treatment. We later relax this assumption to account for heterogeneity in effort

within a treatment. We design our treatments for the experiment by varying the incentives and behavioral motivators

for the workers to identify the parameters of the above model.

7

2.1 Piece Rate Treatments

In the piece rate treatments, each worker works on a task at a given piece rate. We assign four different piece rates

between 0 and 10 cents. In the first treatment, workers are paid no piece rate. In the next three treatments, the piece

rates are 3, 6 and 9 cents per unit of effort. From the M-Turk standards, this variation in piece rates is substantial and

is likely to elicit significant change in effort. The piece rates are earned on top of $1 fixed participation fee. These four

treatments differ only in the piece rates, everything else remains exactly the same.

These piece rates provide evidence on the responsiveness of effort to incentives for our particular task and hence,

allow us to estimate the baseline parameters (s, k, and γ) which will be used to estimate other behavioral parameters.

Formally, in the piece rate treatments, worker observes the piece rate p and then chooses effort e j by maximizing 1.

There is no employer j in these treatments which implies that for any worker, α j = 0 and there is no gift from the

employer implying 1Gi f t = 0. The equilibrium efforts e?j in these treatments is thus given as:

e?p = c′−1 (s+ p) forp ∈ {0,3,6,9}

The solution of effort has one behavioral unknown (s), and two unknowns from the cost function (k and γ). To

back out these parameters, we use effort corresponding to three different piece rates which gives us three equations to

identify these three parameters.

2.2 Altruism Treatments

In the altruism treatments, each worker is matched to an employer (truthfully) and he/she observes the (true) value of

the effort to the matched employer. Each employer earns 1 cent for every 100 points scored by the matched worker. We

set the piece rate to 0. There are three treatments under altruism treatments. In the first treatment (altruism baseline) a

worker does not observe the identity of the matched employer. In the ’altruism black’ and ’altruism white’ treatment

the matched employer is black and white, respectively. Note that our notion of altruism captures “pure altruism” as

well as “warm glow” of the workers, we don’t aim to disentangle between these two.9

Formally, in the altruism treatments, worker i observes the zero piece rate (p = 0), the value of the unit of effort to

9See DellaVigna et al. (2016) for more on the distinction between pure altruism and warm glow.

8

the employer j (v = 0.01), and then chooses effort e j by maximizing 1. There is no gift from the employer implying

1Gi f t = 0. The equilibrium efforts e?j in these treatments is thus given as;

e?j = c′−1 (s+α jv) for j ∈ {Neutral,Black,White}

We are implicitly assuming that altruism parameter can be different for each employer’s group identity. αWhite >

αBlack (αWhite <αBlack) represents the higher (lower) effort for whites as compared to blacks because of greater (lesser)

altruistic preference towards the former as compared to latter. In other words, the difference in effort between the

treatments ‘altruism white’ and ‘altruism black’ is interpreted as resulting from the differential altruistic preferences of

the workers. The three altruism treatments help us identify αNeutral , αBlack, and αWhite, given the baseline parameters.

2.3 Reciprocity Treatments

Reciprocity treatments build on from the altruism treatments and add a positive monetary gift (20 cents) from the

employer to the worker. Rest of the details are exactly the same as in altruism treatments. Thus, the equilibrium effort

is given as;

e?j = c′−1 (s+α jv+ρ j) for j ∈ {Neutral,Black,White}

Once again we are implicitly assuming that reciprocity parameter can be different for each employer’s group

identity. ρWhite > ρBlack (ρWhite < ρBlack) represents the higher (lower) effort for whites as compared to blacks because

of greater (lesser) reciprocal preference towards the former as compared to latter. In other words, the difference in

effort between the treatments ‘reciprocity white’ and ‘reciprocity black’ is interpreted as resulting from the differential

reciprocal preferences of the workers. The three reciprocity treatments help us identify ρNeutral , ρBlack, and ρWhite

given the baseline parameters and the altruism parameters.

9

3 Experiment Design

The main goal of this study is to investigate the possibility of discrimination from the worker side in an online labor

market. We designed the experiment to allow for the possibility of discrimination in the effort by workers towards

the employers. Our experiment is carefully designed to ensure that observed difference in the effort towards the em-

ployers could only realize because of the social preferences of workers, i.e., we shut down the possibility of statistical

discrimination (the employers did not get to make any strategic choices such as wage offer, minutes of work, etc.).

3.1 Task

We design this experiment to observe whether workers discriminate in their effort when working for different employer

types and then to back out the behavioral parameters of distaste. For this purpose, we need a task which is costly to

workers and is not meaningful to any particular race. We settled on a button-pressing task as in Dellavigna and Pope

(2018). The task involves alternating presses of “a” and “b” on a keyboard for 10 minutes. We settled on this task

because it is simple to understand and has features that parallel clerical jobs: it involves repetition, and it gets tiring,

and thus, it tests the motivation of the workers.

3.2 Race Revelation

We take the approach of revealing race via the revelation of skin-color (Doleac & Stein, 2013). To that end, we record

videos of employers while they read off a script explaining and demonstrating the task. The camera placement only

captures the hand of the employer along with the movement of the fingers alternating ‘a’ and ‘b’ button presses. Other

identifiers, such as the face, are not captured in the video to avoid psychological confounds which are associated

with faces, such as attractiveness and trustworthiness (Eckel & Petrie, 2011). The employer’s hand is bare or covered

(with full sleeves and latex gloves) depending on the assigned treatment. For black employers, we only restrict to

participants with dark skin tone to avoid any ambiguity about the race of the person. We use the digital voice for the

videos in the piece rate treatments. We program each video to play with subtitles to aid easier understanding of the

instructions. Here is the link to the sample video corresponding to the ‘Piece Rate - 0 cents’ treatment.

10

3.3 Experiment Flow

The experiment proceeds as follows: (1) We recruit employers from Iowa State University and record videos of them

explaining the task, 2) we then post a HIT on Amazon’s Mechanical Turk for a screener survey for the recruitment of

workers, (3) we invite subjects who consent and meet the recruitment criteria (undisclosed) to initiate the experiment,

(4) upon initiation we assign each subject to one of the treatment groups. Following Czibor, Jimenez-Gomez, and

List (2019), we use the blocked randomization design to assign subjects to treatments. We define blocks based on

demographic information collected in the screener survey (Gender, Age, Race, Education, Income, Political Party

Affiliation, and the Most-lived US State),10 (4) we present instructions to each subject in a pre-recorded video (based

on the assigned treatment). We program our study to require each worker to watch a video, (5) we then elicit beliefs

of each worker about the matched employer, and finally 6) workers work on the task for 10-minutes.

3.3.1 Piece Rate Treatments

In the piece rate treatments, each worker sees a video demonstrating a task with a script: “On the next page, you

will play a simple button-pressing task. The object of the task is to alternately press the ‘a’ and ‘b’ buttons on your

keyboard as quickly as possible for ten minutes. Every time you successfully press the ‘a’ and then the ‘b’ button, you

will receive a point. Note that points will only be rewarded when you alternate button pushes: just pressing the ‘a’ or

‘b’ button without alternating between the two will not result in points. Buttons must be pressed by hand only (key-

bindings or automated button-pushing programs/scripts cannot be used), or task will not be approved. Feel free to

score as many points as you can.” Then the last line is based on the assigned treatment (0, 3, 6 or 9 cents). The wording

is provided in table 1. Even though piece rates are framed in units of 100 points, workers are paid continuously for

each point scored and they are able to see the earned bonus in real time as they score points.

3.3.2 Social Preference Treatments

In the altruism and reciprocity treatments, each video starts with the introduction by the employer: “Hi, I am another

participant in this study who is matched to you. In this study, you will work on a simple button-pressing task, and I

will earn some money depending on how well you do on the task.” Then the script follows the same instructions as in

10See Cavaille (2018) for instructions on implementing sequential blocked randomization for online experiments.

11

Table 1: Summary of treatmentsCategory Treatment Wording Voice Skin Color Sample Video

(1) (2) (3) (4) (5)

Piece Rate Your score will not affect your payment in any way. Muted Concealed Link

As a bonus, you will be paid an extra 3 cents for every 100

points that you score.

Muted Concealed Link

As a bonus, you will be paid an extra 6 cents for every 100

points that you score.

Muted Concealed Link

As a bonus, you will be paid an extra 9 cents for every 100

points that you score.

Muted Concealed Link

Altruism I will earn 1 cent for every 100 points that you score. Your

score will not affect your payment in any way.

Muted Concealed Link

I will earn 1 cent for every 100 points that you score. Your

score will not affect your payment in any way.

Black Black Link

I will earn 1 cent for every 100 points that you score. Your

score will not affect your payment in any way.

White White Link

Reciprocity I will earn 1 cent for every 100 points that you score. In

appreciation to you for performing this task, I have decided

to pay you extra 20 cents as a bonus. Your score will not

affect your payment in any way.

Muted Concealed Link

I will earn 1 cent for every 100 points that you score. In

appreciation to you for performing this task, I have decided

to pay you extra 20 cents as a bonus. Your score will not

affect your payment in any way.

Black Black Link

I will earn 1 cent for every 100 points that you score. In

appreciation to you for performing this task, I have decided

to pay you extra 20 cents as a bonus. Your score will not

affect your payment in any way.

White White Link

Notes: The table list all the treatments in this study. Each piece rate treatment differs just in the last line of thescript, uses no audio, and conceals the skin color of the hand. Social preference treatments (altruism and reciprocity)begin with the introduction of the employer (in the first person), explain the task using the same script as in piece ratetreatments and then differ only in the last paragraph of the script. Both altruism and reciprocity categories have threetreatments, each with black, white, and concealed skin tone of the employer (using gloves). In the social preferencetreatments of concealed skin tone, the ratio of black and white employers will be 1:1.

12

piece rate treatments with the last paragraph being the only difference. The wording is provided in table 1. There are

three treatments each in the category of altruism and reciprocity. Altruism baseline and reciprocity baseline conceals

the skin color of the employer using latex gloves. Although the voice in the baseline treatments can reveal some racial

markers, we recruit an equal number of black and white employers in these treatments, therefore, on average, the

effect of race from voice, if any, should cancel out in the baseline treatments. The videos shown in the altruism black

(white), and reciprocity black (white) reveal the black (white) skin of the employers.

3.4 Recruitment of Subjects

Since this study involves the understanding of worker’s motivation towards their employers, we need to recruit both

employers and workers for this experiment.

3.4.1 Recruitment of Employers

To recruit employers, we invite male student subjects , above the age of 18, from Iowa State University, who racially

identify as either African American or Caucasian. We restrict to male employer subjects to avoid confounds from

identity effects of gender. Holding the sample size fixed, restricting to one social identity give us more power to

make credible inferences. We include employer subjects of only Black and White races because we intend to study

racial discrimination in the context of the United States and also because our race revelation mechanism (as explained

above) works best for these two races. We recruit subjects from Iowa State University for logistical convenience. We

randomly assign each employer subject to one of the six social preference treatments. Based on the assigned treatment,

subjects read from the script and demonstrate the task on a video. Each subject is paid $5 for showing up to our lab

and an additional variable amount depending on the matched workers’ performance. Our final sample include six

employers in each social preference treatment (36 employers).

3.4.2 Recruitment of Workers

We recruit the workers for this experiment from an online labor market, Amazon’s Mechanical Turk. Mechanical Turk

is a crowd-sourcing web-service that allows employers (called requester) to get tasks (called Human Intelligence Tasks

(HITs)) executed by employees (called workers) in exchange for a wage (called reward). Mechanical Turk is a widely

13

used platform in research in economics and gives access to a large pool of applicants at a much affordable rate, thereby,

allowing researchers to conduct a well-powered study. See Paolacci, Chandler, and Ipeirotis (2010) and Paolacci and

Chandler (2014) for a discussion on demographic characteristics and representation of subjects from Mechanical Turk.

To recruit subjects we post a screener survey as the HIT on Mechanical Turk with the following description “Fill

out this 2-minute screener survey to qualify for a study that starts immediately, take up to 15 minutes, and pays

participation bonus $1 with scope to earn extra. You will be required to watch and listen to a video. Do NOT take this

study on mobile.”. Based on the responses in the screener survey; we invite participants above the age of 18 who report

their race as “White or Caucasian” to participate in the experiment. Based on our pilot for this study, we found that it

is difficult to recruit a representative number of Black workers from M-Turk to make a credible inference. Therefore,

for this study we restrict to only white workers and study their effort choices for Blacks versus White employers. We

pay five cents to each subject who do not meet the inclusion criteria for filling out the screener survey.

3.5 Pre-registration

We pre-registered the design on AEA RCT registry as AEARCTR-0003885. Since our choice of the task is the same

as Dellavigna and Pope (2018), we can use results from their study to determine the sample size that can achieve

sufficient power for our study. Dellavigna and Pope (2018) found that the points scored across all treatments have

a standard deviation of around 660 points. Assuming this standard deviation for each treatment and assuming a

minimum detectable effect of 0.2 standard deviations between two treatments, we will need around 400 subjects in

each treatment to have a power of 80 percent. This implies that we will need 400 x 10= 4,000 observations in total

for all ten treatments. We pre-registered the rule for sample size: we aim to recruit 6,000 participants from the United

States within the first three weeks of posting the experiment.

4 Data

4.1 Employers

We recruit employer subjects for this experiment from the Iowa State University (ISU). We invite 50 student subjects

who report their race as black or white to participate in the experiment. Our final sample include 36 employers (18

14

blacks and 18 whites). Each employer subject fill out a short demographic survey upon showing up to our lab for

video recording and are then randomly assigned to one of the six social preference treatments. The demographic

characteristics of the employer subjects in each treatment are presented in table C1.

4.1.1 Pre-Testing of Videos

To test whether the videos correctly reveal the race of the employer, we test these videos using an independent sample

of subjects from the Academic Prolific, a data collection platform. We use Academic Prolific for video testing, rather

than Mechanical Turk, to ensure that M-Turk subjects don’t potentially watch these videos before they participate in

the actual experiment. We invite white subjects from the United States to evaluate these videos. Each subject get

to evaluate one randomly selected video. See figure 1 for the graphical representation of average perception of race

across treatments. Overall the race is correctly perceived more than 80 percent of the times for all the race salient

treatments indicating that our race revelation mechanism works. For the race neutral treatments, only less than 30

percent of the people could guess the race and that is probably the result of random guessing. The pairwise compar-

isons of race perception among these treatments is presented in table C2. The results suggest that the race neutral

treatments (altruism and reciprocity) are statistically indistinguishable from each other and significantly different from

race salient treatments. The perception of race in the treatments ‘Altruism Black’ and ‘Altruism White’ is statistically

indistinguishable, however ‘Reciprocity Black’ is not perceived as accurately as ‘Reciprocity White’.

15

Figure 1: Race PerceptionNotes: This figure shows the proportion of individuals who were able to correctly guess the race of the employer afterwatching a video.

We also evaluate the videos in race salient treatments for perception of skin color, the results are presented in

figure 2. Overall, blacks’ skin is correctly perceived as of darker tone and whites’ as of lighter tone. The pairwise

comparisons of skin color perception among these treatments is presented in table C2. The results suggest that the

black treatments are statistically indistinguishable from each other and significantly different from white treatments.

16

Figure 2: Skin Color PerceptionNotes: This figure shows the average perceived skin tone accross the race salient treatments.

Finally, to ensure that these videos are not perceived differently on soft personality traits such as friendliness,

professionalism, clarity etc., we get these videos rated on those traits. The results for positive traits are presented in

figure B.1 of appendix, and results for negative traits are presented in figure B.2 of appendix. Pairwise comparisons of

means across all the social preference treatments suggest that only the reciprocity black treatment is perceived to be

significantly higher on positive traits while all the other treatments are statistically indistinguishable from each other

on both positive and negative traits (see table C3 of appendix).

4.2 Workers

We ran the experiment for workers for three months from August 2019 to October 2019. Our data collection lasted

longer than expected because of the slow take up of the study on M-Turk. As per the pre-registration, we apply

the following restrictions to the collected data (1) we drop 17 workers for scoring above 4,000 points since this is

physically impossible in 10 minutes and these users may have used some automated script to score points or they may

17

have experienced some technical problem11, 2) we drop 64 workers who score zero points as this may reflect some

malfunction or technical problem in points recording, (3) we drop 4 observations of the workers for participating in the

study more than once as these people have already participated in the study under a potentially different treatment12.

In addition to above, we dropped two observations of workers who could somehow take this study from outside the

United States.13 The final sample consists of 5,945 workers.

The summary statistics of the final sample is presented in table 2. Our sample over represents women, young,

educated, middle-income and democrats as compared to the US population. This is typical of the population on

online platforms. We present results of productivity by the demographics in table C13. Overall in our sample, men

and younger workers are more productive than women and older workers respectively. We present test of balance of

demographic variables across ten treatments in table C10, C11, and C12 of appendix. The treatments are balanced on

all the observed variables, this is no surprise since we use blocked randomization to assign subjects to the treatments.

Since the workers characteristics are balanced across treatments, there is no reason to believe that more/less productive

workers are assigned to a specific treatment.

11We informed each worker to not to use any automated scripts/programs during the instructions.12A worker can participate in our study only once, these exceptions must be an error on part of M-Turk.13The study was restricted to workers from the United States only, but these people must have used some proxy or VPN to be able to take the

study. We could identify them from the GPS coordinates recorded by Qualtrics.

18

Table 2: Summary Statistics, Worker Sample

(1) (2)Sample US Population

GenderFemale 0.58 0.50Male 0.41 0.48

RaceWhite or Caucasian 1.00 0.77

Age18-24 0.12 0.1325-30 0.38 0.1131-40 0.26 0.1741-50 0.14 0.1951-64 0.08 0.2365 and over 0.03 0.17

EducationLess than high school 0.01 0.13High school or equivalent 0.13 0.30Some college 0.28 0.33College graduate 0.41 0.20Graduate or professional degree 0.18 0.12

IncomeLess than $20,000 0.17 0.20$20,000 - $44,999 0.31 0.26$45,000 - $99,999 0.38 0.33$100,000 - $149,999 0.09 0.12$150,000+ 0.03 0.08

Political AffiliationDemocrat 0.39 0.31Independent 0.28 0.38Republican 0.27 0.29

Most lived US StateBlue 0.31 0.47Red 0.20 0.14Swing 0.49 0.39

Observations 5945 234564071Notes: The table presents demographic information of worker subjects. Column (1) presents proportion of the workersubjects by their gender, race, age, education, income, party, and the most lived state in the United States. Column(2) presents these demographics for US population (above 18) based on 2010 Census. Estimates of population bypolitical affiliation and by blue, red, and swing state are based on Gallup polling survey 2019.

5 Results

We present the average effort by workers against each treatment in column 1 of table 3 and in figure 3. Overall the

incentives have a strong effect on effort, raising performance from 1627 points (no piece rate) to 2060 points (3-cent

piece rate) and 2127 points (9-cent piece rate). The standard error for the mean effort per treatment is around 30 points

19

or less, implying that differences across treatments larger than 85 points are statistically significant.

There is statistically significant evidence for altruism. Workers put more effort when there work benefit the em-

ployer (altruism neutral treatment) as compared to the piece rate 0-cent treatment. The one cent return to the employer

induces an effort of 1746 points as compared to 1627 points in the 0-cent treatment. Interestingly, workers are signif-

icantly more altruistic towards black employers than white employers. The effort corresponding to black employers

is 1798 points, which is significantly higher than the effort provided to the white employers (1708 points). However,

effort towards any employer race is not significantly different from the effort towards the neutral employer, implying

very week evidence for any sort of discrimination in altruistic preferences.

In the reciprocity treatments, worker receive an unexpected gift of 20-cents from the employer, unconditional on

performance. This positive gift from the employer, does not induce a significant increase in effort as compared to

the altruism treatment (1771 points in reciprocity neutral treatment as compared to 1746 points in altruism neutral

treatment). The reciprocal response to the employer’s racial identity is also insignificant, implying that, on average,

the workers do not reciprocate towards any employer race. This result is consistent with the literature which finds

weak evidence for positive reciprocity (such as Kube, Maréchal, and Puppe (2006)).

Although our treatments are balanced on the observed worker variables and the employer characteristics, we

present the regression results (for robustness sake) by controlling for these variables in the table C8. We observe

that workers pro-altruistic response becomes less significant after controlling for the demographic variables and em-

ployer fixed effects, while the reciprocal response also stay statistically indifferent from zero after controlling for the

demographics and employer fixed effects.

In column 2 of table 3, we restrict to workers who could correctly perceive the race of the employer as black, white

or neutral in the social preference treatments. Most of the workers (88 percent) were able to correctly perceive the

race of the employer in the race salient treatments however only few workers could remain indecisive about the race

of the employer in the race neutral treatments. More than half (five percent) of the workers in race neutral treatments

believed the employer to be white (black), even though there were no racial markers in the video that could identify

the race of the employer. Restricting to workers who could correctly perceive the employer race do not substantially

effect the direction or magnitude of results.

20

Table 3: Effort by Treatment

(1) (2)All Workers Correctly Perceived Race

N Mean (s.e) N Mean (s.e)Piece Rate - 0 cents 599 1627.07 599 1627.07

(28.56) (28.56)Piece Rate - 3 cents 595 2059.83 595 2059.83

(24.19) (24.19)Piece Rate - 6 cents 592 2046.68 592 2046.68

(23.62) (23.62)Piece Rate - 9 cents 588 2127.37 588 2127.37

(23.01) (23.01)Altruism - Neutral 591 1746.06 261 1724.87

(29.15) (43.70)Altruism - Black 601 1798.37 494 1807.68

(27.55) (29.58)Altruism - White 592 1708.09 557 1715.24

(28.90) (29.52)Reciprocity - Neutral 608 1771.15 265 1766.99

(27.95) (41.63)Reciprocity - Black 590 1803.61 470 1818.78

(26.95) (29.73)Reciprocity - White 589 1798.23 561 1803.75

(29.58) (30.33)Total 5945 1848.08 4982 1865.98

(8.80) (9.49)Notes: The table presents the effort choices in each treatment. Column 1 reports the effort choices by all the workers,column 2 reports the effort choices by workers who were able to correctly perceive the race of the employer as neutral,black or white in social preference treatments.

21

Figure 3: Effort by Treatment - All WorkersNotes: This figure presents the average score and confidence interval for each of ten treatments for all workers. Eachtreatment has about 590 participants.

5.1 Distribution of Effort

Beyond the average effort, we present the distribution of effort in all the treatments in figure B.3 of appendix and by

each treatment in figure B.4 of appendix. Overall, very few workers score below 500 points and even fewer score

above 3000 points.

Figure 4a presents the cumulative distribution function for the piece rate treatments. Incentives induce a clear

rightward shift in effort relative to 0-cent treatment. However, there is not much change in effort between 3-cent and

6-cent treatments. Figure 4b shows the strong evidence for altruistic preferences as observed by the clear rightward

shift of the altruism treatment as compared to the 0-cent treatment. The reciprocity treatment is indistinguishable from

the altruism treatment, implying a lack of reciprocal preferences. Figure 4c shows that altruism is stronger towards

blacks as compared to whites while the cumulative density function is indistinguishable for reciprocity-black and

reciprocity-white treatments.

22

(a)

(b)

(c)

Figure 4: Cumulative distribution functionNotes: The figure presents the cumulative distribution function of points for the workers in each of the treatmentsfeatured. The sample size in each treatment is approximately 590 subjects. Figure a features the four piece ratetreatments (no piece rate, 3-cent per 100 points, 6 cents per 100 points, and 9 cents per 100 points). Figure b presentsthe results for the race-neutral treatments. Figure c presents the results for the race-salient treatments.23

5.2 Evolution of Effort

We present the evolution of effort over the 10-minutes period in figure 5. Figure 5b shows that in the social preference

treatments, overall the effort declines over time, this may be due to workers getting tired as they continue working over

the 10-minutes interval. However, the piece rate treatments are able to sustain the consistent higher effort throughout

the time interval (figure 5a), with workers in 9-cent treatment pushing extra hard near the end.

24

(a)

(b)

Figure 5: Evolution of effort over timeNotes: This figure presents the effort over time for selected treatments. The y axis indicates the average number ofpoints scored in that treatment per minute.

25

5.3 Heterogeneity

5.3.1 Heterogeneity by Demographics

To examine the heterogeneity in our average treatment effects based on demographic characteristics of the sample,

we present the conditional average treatment effects in table C4 and C5 of appendix for altruism treatments and

in table C6 and C7 of appendix for reciprocity treatments. Overall, we do not find evidence of heterogeneity on

the basis of gender, age, education, income, and state voting pattern for both altruism and reciprocity treatments.

However, we do find evidence for heterogeneity in altruism on the basis of party affiliation. Interestingly, we find that

republicans and independents exert significantly more effort than democrats for the black employers relative to the

white employers, indicating the presence of pro-black altruistic preferences among republicans and independents as

compared to democrats.

5.3.2 Heterogeneity by the share of black population in the neighborhood

Following Andreoni, Payne, Smith, and Karp (2016), we explore the effects of local racial composition on social

preferences of the workers in our sample. We condition on the zip code level racial composition of the worker, and

examine the difference in effort provided to black versus white employers. Figure 6 presents the conditional average

treatment effects for each decile of the share of black population for workers who correctly perceived the employer

race. Overall, the difference in effort provided to the black versus white employers is statically zero at each level of

black share of population.

5.3.3 Heterogeneity by Geographical Area

It is a well established fact that racial disparities are not equally distributed across the United States. We present the

summary of workers performance by their geographical area in table 4. Interestingly, there is a week evidence in favor

of workers from south being relatively pro-social to black employers.

5.3.4 Heterogeneity by Implicit Biases

We examine the heterogeneity in treatment effects based on the implicit biases of the workers as measured by the

implicit association test (IAT). IATs are widely used in social psychology to measure implicit and unconscious biases

26

Figure 6: Treatment effects by local racial composition

Notes: The figure presents the conditional average treatment effects (conditioned on the share of blacks in a zip code).The x-axis represents deciles of the share of black population in a zip code. Measure of conditional treatment effect isobtained by pooling data from race salient social preference treatments of workers who could correctly perceive theemployer race and running a regression of Points on Employer Race for each decile of the black share. The cutoffvalues of the deciles are 0, 0.009, 0.018, 0.03, 0.045, 0.065, 0.094, 0.137, 0.207, and 0.351.

Table 4: Heterogeneity by Geographical Area

Regions(1) (2) (3) (4)

North East Mid West South WestWhite or Caucasian -38.42 -38.60 -79.37 -27.42

(71.19) (57.03) (49.29) (71.46)Constant 1750.3 1859.5 1834.1 1765.5

(51.20) (40.59) (36.38) (54.43)Observations 377 529 771 405

Notes: The table presents the conditional average treatment effect by the geographical location of the worker. Measureof conditional treatment effect is obtained by pooling data from race salient social preference treatments of workerswho could correctly perceive the employer race and running a regression of Points on Employer Race for each geo-graphical region. Standard errors in parenthesis.

27

towards a particular group. The test involves categorizing two sets of words to the left or right hand side of the

computer screen. Implicit bias is measured by a time difference in associating good or bad words to the relevant group

identities. The idea is that making a response is easier when closely related items share the categorization to the same

side of the screen. In case of race IAT, we would say that one has an implicit preference for white people relative to

black people if they are faster to categorize words when white face and good words (friend, glorious, enjoy, joyous,

terrific, beautiful, magnificent, and fabulous) share a response key and black faces and bad words (detest, poison,

nasty, disgust, pain, despise, sadness, evil) share a response key, relative to the reverse. The race IAT data by project

implicit contains historical record of tests taken on the project’s website. These tests can be taken by anyone from

anywhere in the world. For our purpose we restrict to white individuals from the United States and examine the data

of more than two million tests takers over the time period from 2006 to 2018.

According to the typical thresholds in literature (Greenwald, Nosek, & Banaji, 2003; Hahn, Judd, Hirsh, & Blair,

2014; Rooth, 2010), IAT scores below -0.15 indicate some preference for minorities; scores between -0.15 and 0.15

indicate little to no bias; scores between 0.15 and 0.35 indicate a slight bias against minorities; and scores above 0.35

show moderate to severe bias against minorities. The average score (standard deviation) of white test takers in our

sample is 0.38 (0.42) implying that on average white people have moderate to severe implicit bias against blacks.

We map the county level IAT score to workers in our sample based on the worker’s geographic location. Our worker

sample comes from 190 counties spanning over all 50 states in the United States. The average (standard deviation) IAT

score for white workers in our counties is 0.39 (0.04). Like black share, we explore the effects of local IAT score on

social preferences of the workers in our sample. We condition on the county level IAT score of the worker, and examine

the difference in effort provided to black versus white employers. Figure 7 presents the conditional average treatment

effects for each decile of the IAT score for workers who could correctly perceive the employer race. Overall, there is

some indication that workers with low implicit bias exhibit higher social preference towards the black employers as

compared to white employers. However, at higher level of implicit biases, the difference in effort provided to the black

versus white employers is statistically zero.

28

Figure 7: Treatment effects by local IAT

Notes: The figure presents the conditional average treatment effects (conditioned on the IAT score of the worker’scounty). The x-axis represents deciles of the IAT score at county level. Measure of treatment effect is obtained bypooling data from race salient social preference treatments of workers who could perceive the employer race correctlyand running a regression of Points on Employer Race for each decile of the IAT score. The cutoff values of the decilesare 0.295, 0.349, 0.376, 0.381, 0.386, 0.396, 0.404, 0.413, 0.415, and 0.444.

29

6 Estimates of Behavioral Parameters

We designed our experiment with the structural model given in Section 2. The advantage of designing field experiments

on the basis of model of behavior is that it allows researchers to estimate the nuisance parameters in the environment

that is relevant to the decision making (DellaVigna, 2017). Because of the simplicity of our task, there are only three

nuisance parameters that we need to estimate. We will thus use data from the piece rate treatments to identify these

parameters. Once we have the estimate of these nuisance parameters, we can estimate behavioral parameters using

data from the social preference treatments. We closely follow the estimation procedure in Dellavigna and Pope (2018)

to estimate our model.

6.1 Minimum Distance Estimation

We first use minimum distance estimation method to estimate these parameters. In the minimum distance estimation

one identifies the set of moments in data (average effort) and then find the set of model parameters that minimizes

the distance between the empirical moments and the theory-predicted moments. To estimate nuisance parameters, we

use data from the piece rate treatments. Specifically, we use the average effort corresponding to three piece rates (0

cents, 3 cents and 9 cents), to estimate γ , s, and k. Specifically, in the case of power cost function, to estimate nuisance

parameters, we use first moments from the piece rate treatments and solve the following equations

ep =1γ[log(s+ p)− log(k)] for p ∈ {0,0.03,0.09}

where ep is the average effort in the piece rate p treatment. These parameters estimates are used to present marginal

cost and marginal benefits curve in figure 8.

Once these parameters are estimated, we use average effort corresponding to altruism neutral, altruism black and

altruism white treatment to estimate behavioral parameters αNeutral , αBlack, and αWhite respectively. Specifically, for

power cost function, we solve the following equations for α j for j ∈ {Neutral,Black,White} taking estimates of γ , s,

and k as given

log(eα j) =1γ[log(s+α jv)− log(k)] for j ∈ {Neutral,Black,White}

30

Figure 8: Illustration of the Model: Marginal Benefits and Cost Curves

Notes: The figure presents the marginal benefit and marginal cost curves using minimum-distance estimates for powercost function.

where eα j is the average effort in the altruism j treatment.

Similarly, to calculate reciprocity parameters for neutral (ρNeutral), black (ρBlack) and white (ρWhite) employers, we

use average effort from reciprocity neutral, reciprocity black, and reciprocity white treatments and solve the following

equations taking estimates of γ , s, k, and α j for j ∈ {Neutral,Black,White} as given

log(eρ j) =1γ[log(s+ρ j +α jv)− log(k)] for j ∈ {Neutral,Black,White}

where eρ j is the average effort in the reciprocity j treatment.

Estimates using exponential cost function are similarly calculated. Table 5 presents the parameter estimates for

power cost function (column 1) and exponential cost function (column 3). The standard errors for these parameter

estimates are estimated using a bootstrap procedure with thousand draws.

6.2 Non-Linear Least Squares Estimation

The minimum distance estimator just relies on the moment and hence, does not use all the variation in the data. There

are methods such as maximum likelihood and non-linear least squares that can be used to estimate these parameters

using all the variation in the data. We use non-linear least square method to estimate these parameters allowing for the

heterogeneous cost of effort. Allowing for a heterogeneous marginal cost of effort in 1, we assume for a worker i, for

31

a power cost case, c(ei j) =ke1+γ

i j1+γ

exp(−γεi j) with εi j ∼ N(0,σ2ε ). The first order condition 4 can then be written as;

s+1Gi f tρ j +α jv+ p− keγ

i jexp(−γεi j) = 0

Taking the last term to the right and taking logs, we obtain

log(s+1Gi f tρ j +α jv+ p)+ εi j = log(k)+ γlog(ei j)− γεi j

Solving for log(ei j), we obtain the estimating equation

log(ei j) =1γ

[log(s+1Gi f tρ j +α jv+ p)− log(k)

]+ εi j (5)

Similarly using exponential cost function, we get

ei j =1γ

[log(s+1Gi f tρ j +α jv+ p)− log(k)

]+ εi j (6)

Equations 5 and 6 can be estimated with non-linear least squares (NLS). Table 5 presents the NLS parameter

estimates for power cost function (column 2) and exponential cost function (column 4). The NLS parameter estimates

are nearly identical to minimum distance estimation for the exponential cost case. The model predictions are also very

similar to minimum distance ones.

The NLS estimates for the power cost function yield a lower curvature than the minimum-distance estimates

(γNLS = 20.29 versus γMD = 34.05). The NLS model matches expected log effort, while the minimum-distance matches

the log of expected effort. Both NLS and minimum-distance fit the in sample moments and make similar predictions

for the 6-cent piece rate treatment.

The altruism parameters are essentially estimated to be zero for all the employers with altruism towards blacks

being marginally higher than altruism towards whites in all four specifications. The reciprocity estimates also indicate

almost no effect from the gift exchange for any employer in all the specifications.

32

Table 5: Parameter Estimates

Power cost of effort Exponential cost of effortMinimum distance NLS on Individual Minimum distance NLS on individual

estimator on average effort effort estimator on average effort effort(1) (2) (3) (4)

Baseline Parameters

Curvature γ of cost of effort function 34.05 20.30 0.0163 0.0163(15.9) (8.85) (.0207) (.00807)

Intrinsic motivation s (cents per point) 0.00000977 0.00000802 0.0000264 0.0000264(.000246) (.000032) (.000389) (.000101)

Level k of cost of effort function 4.50e-115 2.98e-70 8.58e-17 8.58e-17(2.7e-46) (2.5e-68) (7.1e-09) (1.5e-15)

Altruism Parameters

Altruism αNeutral towards neutral employer 0.00983 0.000426 0.0156 0.0156(.00779) (.0017) (.0103) (.0427)

Altruism αBlack towards black employer 0.0285 0.000776 0.0402 0.0402(.0186) (.00274) (.0226) (.0953)

Altruism αWhite towards white employer 0.00413 0.000270 0.00722 0.00722(.00367) (.00129) (.00552) (.0215)

Reciprocity Parameters

Reciprocity ρNeutral towards neutral employer 0.0000676 0.0000272 0.0000921 0.00124(.000136) (.000103) (.000173) (.00318)

Reciprocity ρBlack towards black employer 0.0000307 0.0000395 0.0000381 0.00220(.000265) (.00014) (.000308) (.00513)

Reciprocity ρWhite towards white employer 0.000243 0.0000255 0.000328 0.00200(.00021) (.0001) (.000257) (.00477)

Implied effort at 6-cents expected log effort(observed effort 2047, log 7.624000000000001) 2102 7.746 2102 2102.4

Notes: This table reports the structural estimates of the model in section 2. Column (1) and (3) use a minimum-distance estimator employing three moments (averageeffort in three piece rate treatments) and three parameters (γ , s and k), and is thus exactly identified. Column (2) and (4) use a non-linear least squares employingindividual effort in all the treatments and thus estimating all the parameters simultaneously. We use power cost (column 1 and 2) and exponential cost (column 3 and4) function to estimate the model. Implied effort is calculated using estimated parameters for each model. For the altruism parameters, the baseline parameters aretaken as given and the average effort for neutral, black, and white employers is used to estimate αNeutral , αBlack , and αWhite from the altruism treatments. Similarly forthe reciprocity parameters, the baseline and altruism parameters are taken as given and the average effort corresponding to reciprocity neutral, reciprocity black, andreciprocity white is used to estimate ρNeutral , ρBlack , and ρWhite. Standard errors for minimum-distance estimator are calculated by taking a bootstrap sample of 1000draws and recalculating these parameters for each draw.

33

7 Conclusion

This paper uses insights from behavioral and experimental economics to shed light on a pressing issue in American

society, namely, racial discrimination. Almost all of the economic literature on discrimination presume that labor

market discrimination goes from employers to workers, this paper instead asks, could it be that workers may also

discriminate against their out-race employers in an online labor market? The results suggest that workers do not

discriminate in effort provision against black employers, instead black employers illicit statistically higher effort from

workers as compared to white employers. There seems to be suggestive evidence that the higher effort towards black

employers is driven by the workers with relatively lower implicit bias against blacks. Worker with higher implicit bias

do not provide disparate effort depending on employer’s race.

It should be noted that our results may not be externally valid to settings in which employer-worker engagement

is longer and involves physical interaction. In those settings, it is likely that workers may still exhibit discrimination

against the employers from a particular group. Future research needs to test worker side discrimination in different

environments.

34

References

Abel, M. (2019). Do Workers Discriminate against Female Bosses?

Akerlof, G. A., & Kranton, R. E. (2005). Identity and the Economics of Organizations. Journal of Economic

Perspectives, 19(1), 9–32.

Andreoni, J., Payne, A. A., Smith, J., & Karp, D. (2016). Diversity and donations: The effect of religious and ethnic

diversity on charitable giving. Journal of Economic Behavior and Organization.

Ayalew, S., Manian, S., & Sheth, K. (2018). Discrimination from Below: Experimental Evidence on Female Leader-

ship in Ethiopia.

Becker, G. S. (1957). The economics of discrimination. University of Chicago Press.

Becker, G. S. (1974). A theory of Social Interactions. The Journal of Political Economy, 82(6), 1063–1093.

Benjamin, D. J., Choi, J. J., & Strickland, A. J. (2010). Social Identity and Preferences. American Economic Review,

100(September), 1913–1928.

Benson, A., Board, S., & Meyer-ter Vehn, M. (2019). Discrimination in Hiring: Evidence from Retail Sales.

Bertrand, M., Chugh, D., & Mullainathan, S. (2005). Implicit Discrimination. American Economic Review, 95(2),

94–98.

Bertrand, M., & Duflo, E. (2016). Field Experiments on Discrimination. Prepared for the Handbook of Field

Experiments.

Bohren, J. A., Haggag, K., Imas, A., & Pope, D. (2019). Inaccurate Statistical Discrimination.

Cavaille, C. (2018). Implementing Blocked Randomization in Online Survey Experiments. , 1–14.

Chakrabortyy, P., & Serra, D. (2019). Gender differences in top leadership roles: Does worker backlash matter?

Charness, G., Rigotti, L., & Rustichini, A. (2007). Individual Behavior and Group Membership. American Economic

Review, 97(4), 1340–1352.

Chen, Y., & Li, S. X. (2009). Group Identity and Social Preferences. The American Economic Review, 99(1), 431–457.

Craig, A. C., & Fryer, R. G. (2018). Complementary Bias : A Model of Two-Sided Statistical Discrimination.

Cuddy, A. J. C., Fiske, S. T., & Glick, P. (2007). The BIAS map: Behaviors from intergroup affect and stereotypes.

Journal of Personality and Social Psychology, 631–648.

Czibor, E., Jimenez-Gomez, D., & List, J. A. (2019). The dozen things experimental economists should do (more of).

35

DellaVigna, S. (2017). Structural Behavioral Economics. Handbook of Behavioral Economics(December).

DellaVigna, S., List, J. A., Malmendier, U., & Rao, G. (2016). Estimating social preferences and gift exchange at

work. NBER Working Paper Series, 53(9), 1689–1699. doi: 10.1017/CBO9781107415324.004

Dellavigna, S., & Pope, D. (2018). What motivates effort? Evidence and expert forecasts. Review of Economic

Studies, 85(2), 1029–1069.

Doleac, J. L., & Stein, L. C. D. (2013). The visible hand: Race and online market outcomes. Economic Journal,

123(572), 469–492.

Donohue, J. J. (2007). Antidiscrimination Law. In Handbook of law and economics (pp. 1387–1472).

Eckel, C. C., & Petrie, R. (2011). Face value. American Economic Review, 101(4), 1497–1513.

Fershtman, C., & Gneezy, U. (2001). Discrimination in a segmented society: An experimental approach. The

Quarterly Journal of Economics, 116(1), 351–377.

Freeman, D. J., Halevy, Y., & Kneeland, T. (2019). Eliciting risk preferences using choice lists. Quantitative

Economics, 10(1), 217–237.

Glover, D., Pallais, A., & Pariente, W. (2017). Discrimination as a Self-Fulfilling Prophecy: Evidence from French

Grocery Stores. Quarterly Journal of Economics, 1219–1260.

Gneezy, U., & List, J. A. (2006). Putting behavioral economics to work: resting for gift exchange in labor markets

using field experiments. Econometrica, 74(5), 1365–1384.

Greenwald, A. G., Nosek, B. A., & Banaji, M. R. (2003). Understanding and Using the Implicit Association Test: I.

An Improved Scoring Algorithm. Journal of Personality and Social Psychology.

Grossman, P. J., Eckel, C. C., Komai, M., & Zhan, W. (2019). It pays to be a man: Rewards for leaders in a

coordination game. Journal of Economic Behavior and Organization, 161, 197–215.

Hahn, A., Judd, C. M., Hirsh, H. K., & Blair, I. V. (2014). Awareness of implicit attitudes. Journal of Experimental

Psychology: General.

Hsieh, C.-T., Hurst, E., Jones, C. I., & Klenow, P. J. (2019). The Allocation of Talent and U.S. Economic Growth.

Econometrica, 87(5), 1439–1474.

Katz, L., & Krueger, A. (2019). Understanding Trends in Alternative Work Arrangements in the United States.

Korn Ferry. (2019). The Black P&L Leader: Insights and Lessons from Senior Black P&L Leaders in Corporate

36

America (Tech. Rep.). Korn Ferry.

Kube, S., Maréchal, M. A., & Puppe, C. (2006). Putting reciprocity to work-positive versus negative responses in the

field. University of St. Gallen Economics Discussion Paper.

Kuziemko, I., Norton, M. I., Saez, E., & Stantcheva, S. (2015). How Elastic Are Preferences for Redistribution?

Evidence from Randomized Survey Experiments. American Economic Review, 105(4), 1478–1508.

Levine, R., & Rubinstein, Y. (2017). Smart and Illicit: Who Becomes an Entrepreneur and Do They Earn More? 2The

Quarterly Journal of Economics, 132(2), 963–1018.

Neuberg, S. L., & Cottrell, C. A. (2006). Evolutionary Bases of Prejudices. Evolution and social psychology, 163–

187.

Neumark, D. (2018). Experimental Research on Labor. Journal of economic literature, 56(3), 799–866.

Paolacci, G., & Chandler, J. (2014). Inside the Turk: Understanding Mechanical Turk as a Participant Pool. Current

Directions in Psychological Science, 23(3), 184–188.

Paolacci, G., Chandler, J., & Ipeirotis, P. (2010). Running experiments on amazon mechanical turk. Judgment and

Decision making, 5(5), 411–419.

Phelps, E. S. (1972). The Statistical theory of Racism and Sexism. American Economic Review, 62(4), 659–661.

Riach, P. A., & Rich, J. (2002). Field experiments of discrimination in the market place. The Economic Journal,

112(483), 480–518.

Rich, J. (2014). What Do Field Experiments of Discrimination in Markets Tell Us? A Meta Analysis of Studies

Conducted since 2000.

Rooth, D.-O. (2010). Automatic associations and discrimination in hiring: Real world evidence. Labour Economics.

Sasaki, M. (1999). An Equilibrium Search Model with Coworker Discrimination. Journal of Labor Economics,

377–407.

Tajfel, H. (1970). Experiments in Intergroup Discrimination. Scientific American, 96–103.

37

A Experiment Material Appendix

38

B Miscellaneous Figures

Figure B.1: Perception of Positive Personality TraitsNotes: This figure presents the box-plot of average of positive traits as rated by the evaluators. After the evaluatorswatched the video they were asked “Please rate the following characteristics about the the person in the above video”.The positive traits were friendliness, confidence, encouragement, trustfulness, clarity, and motivation .

39

Figure B.2: Perception of Negative Personality TraitsNotes: This figure presents the box-plot of average rating of negative traits by the evaluators. After the evaluatorswatched the video they were asked “Please rate the following characteristics about the person in the above video”.The negative traits were arrogance, laziness, bossinesss, rudeness, hostility, and undermining.

40

Figure B.3: Distribution of effortNotes: This figure plots a histogram of the observed points over all 10 treatments.

41

Figure B.4: Distribution of effort by TreatmentNotes: This figure plots a histogram of the observed points by each of the 10 treatments.

42

C Miscellaneous Tables

Table C1: Demographic Information of Employer Subjects

(1) (2) (3)All Subjects Blacks Whites

GenderMale 1.00 1.00 1.00

Female 0.00 0.00 0.00

RaceBlack or African American 0.50 1.00 0.00

White or Caucasian 0.50 0.00 1.00

Age18-24 0.78 0.61 0.94

25-34 0.14 0.22 0.06

35-44 0.06 0.11 0.00

45-54 0.03 0.06 0.00

EducationHigh school or equivalent 0.06 0.00 0.11

Some college 0.64 0.50 0.78

College graduate 0.19 0.28 0.11

Master’s degree 0.08 0.17 0.00

Doctoral degree 0.03 0.06 0.00

Most lived stateBlue 0.28 0.22 0.33

Red 0.03 0.06 0.00

Swing 0.69 0.72 0.67

Observations 36 18 18Notes: The table presents demographic information of employer subjects. Column (1) presents proportion of all theemployer subjects by their gender, race, age and education. Column (2) and column (3) presents these information foronly black and white employers respectively.

43

Table C2: Test of Difference of Perception of Race and Skin ColorPanel A: Average Perception of Race

(1)Race Perception

Proportion SE GroupAltruism 0.29 (0.03) 1Altruism Black 0.91 (0.03) 23Altruism White 0.98 (0.03) 3Reciprocity 0.26 (0.03) 1Reciprocity Black 0.83 (0.03) 2Reciprocity White 0.96 (0.03) 3Degrees of Freedom 1016

Panel B: Average Perception of Skin Color

(1)Skin Color Perception

Mean SE GroupAltruism Black 4.81 (0.05) 1Altruism White 2.05 (0.05) 2Reciprocity Black 4.57 (0.05) 1Reciprocity White 2.11 (0.05) 2Degrees of Freedom 667

Notes: Panel A presents the proportion of subjects who could correctly guess the race of the employer in the video.Panel B presents the average skin color as perceived by the subjects in each treatment. The skin color can vary from 1to 6 where 1 represents the ‘light, pale white’ while 6 represents the ‘very dark brown to black’ skin tone. Proportionssharing a digit in the ‘Group’ column are not significantly different at the 5% level. The comparisonwise error rate isadjusted using the Bonferroni method.

Table C3: Test of Difference of Personality Traits

(1) (2)Positive Traits Negative Traits

Mean SE Group Mean SE GroupAltruism 3.27 (0.07) 12 1.33 (0.04) 12Altruism Black 3.33 (0.07) 12 1.19 (0.04) 1Altruism White 3.15 (0.07) 1 1.30 (0.04) 12Reciprocity 3.26 (0.07) 12 1.38 (0.04) 2Reciprocity Black 3.51 (0.07) 2 1.24 (0.04) 12Reciprocity White 3.28 (0.07) 12 1.28 (0.04) 12Degrees of Freedom 852 929

Notes: The table presents the average of perceived positive and negative traits across the social preference treatments.The perception of the trait can vary from 1-Not at all to 5-Extremely. Positive Trait is constructed by taking an averageof the ratings on; friendliness, confidence, encouragement, trustfulness, clarity, and motivation . Negative Trait is con-structed by taking an average of the ratings on; arrogance, laziness, bossinesss, rudeness, hostility, and undermining.Means sharing a digit in the group label are not significantly different at the 5% level. The comparisonwise error rateis adjusted using the Bonferroni method.

44

Table C10: Balance Checks - Piece Rate Treatments

(1) (2) (3) (4)Piece Rate - 0 Piece Rate - 3 Piece Rate - 6 Piece Rate - 9

GenderFemale -0.01 (0.01) 0.01 (0.01) -0.00 (0.01) 0.01 (0.01)

Age25-30 0.01 (0.01) 0.01 (0.01) 0.02 (0.01) 0.01 (0.01)31-40 0.01 (0.01) 0.00 (0.01) 0.03 (0.01) -0.00 (0.01)41-50 0.01 (0.02) 0.00 (0.02) 0.02 (0.02) 0.01 (0.02)51-64 -0.00 (0.02) -0.00 (0.02) 0.04 (0.02) 0.01 (0.02)65 and over 0.01 (0.03) 0.02 (0.03) 0.01 (0.03) 0.01 (0.03)

EducationHigh school or equivalent 0.01 (0.05) 0.06 (0.05) -0.03 (0.05) -0.05 (0.05)Some college 0.02 (0.05) 0.03 (0.05) -0.02 (0.05) -0.05 (0.05)College graduate 0.02 (0.05) 0.03 (0.05) -0.03 (0.05) -0.06 (0.05)Graduate or professional degree 0.01 (0.05) 0.04 (0.05) -0.02 (0.05) -0.05 (0.05)

Income$20,000 - $44,999 -0.02 (0.01) -0.00 (0.01) -0.00 (0.01) -0.01 (0.01)$45,000 - $99,999 -0.02 (0.01) -0.00 (0.01) -0.00 (0.01) -0.01 (0.01)$100,000 - $149,999 -0.04 (0.02) -0.00 (0.02) 0.01 (0.02) 0.01 (0.02)$150,000+ -0.05 (0.02) 0.02 (0.02) 0.04 (0.02) 0.00 (0.02)

Most lived US StateBlue 0.01 (0.01) 0.01 (0.01) -0.00 (0.01) 0.00 (0.01)Red 0.02 (0.01) -0.01 (0.01) 0.01 (0.01) 0.00 (0.01)

PartyDemocrat 0.00 (0.01) -0.01 (0.01) 0.01 (0.01) 0.01 (0.01)Republican 0.01 (0.01) -0.01 (0.01) 0.02 (0.01) -0.00 (0.01)

Constant 0.09 (0.05) 0.06 (0.05) 0.09 (0.05) 0.14 (0.05)Observations 5945 5945 5945 5945R2 0.002 0.003 0.004 0.002F 0.78 0.87 1.24 0.65

Notes: In Columns 1-4 we regress the assignment to different conditions (Piece Rate Treatments) on the workerdemographics. Standard errors in parentheses.

45

Table C4: Heterogeneous Treatment Effects - Altruism Treatments

(1) (2) (3)Gender Age Education

White or Caucasian -60.32 16.31 -128.2(52.19) (116.6) (115.2)

Male 154.6(56.95)

White or Caucasian ×Male -74.43(80.90)

Age 25 - 34 35.29(90.83)

Age 35 - 44 63.93(95.68)

Age 45 - 54 -23.21(111.1)

Age 55 - 64 -243.6(129.3)

Age 65 or older -46.72(194.5)

White or Caucasian × Age 25 - 34 -125.9(133.2)

White or Caucasian × Age 35 - 44 -152.5(140.8)

White or Caucasian × Age 45 - 54 -96.19(158.2)

White or Caucasian × Age 55 - 64 46.03(184.5)

White or Caucasian × Age 65 or older -275.0(281.4)

Some college -130.2(95.29)

College graduate -198.7(87.83)

Graduate or professional degree -160.0(100.2)

White or Caucasian × Some college 127.9(139.2)

White or Caucasian × College graduate 3.194(129.7)

White or Caucasian × Graduate or professional degree 36.75(149.4)

Constant 1735.0 1790.5 1946.4(36.82) (78.53) (76.77)

Observations 1187 1192 1193Notes: The table presents the estimates from an OLS regression of Points in the race salient altruism treatments onthe employer’s race. The omitted employer is the Black employer. Column (1), (2), and (3) test for the heterogeneityin treatment effects by gender, age, and education respectively. The omitted categories for gender, age, and educationare female, age between 18 and 24, and high school or less. Standard errors in parentheses.

46

Table C5: Heterogeneous Treatment Effects - Altruism Treatments

(1) (2) (3)Income Political Affiliation State Voting Pattern

White or Caucasian 26.40 -156.7 -32.94(98.09) (71.10) (72.13)

Income $20,000 - $44,999 31.32(85.53)

Income $45,000 - $99,999 42.77(83.30)

Income $100,000 - $149,999 1.883(119.2)

Income $150,000+ 86.45(177.0)

White or Caucasian × Income $20,000 - $44,999 -185.3(121.3)

White or Caucasian × Income $45,000 - $99,999 -110.1(117.6)

White or Caucasian × Income $100,000 - $149,999 -33.72(165.7)

White or Caucasian × Income $150,000+ -190.2(250.1)

Democrat -109.3(66.71)

Republican 31.42(73.48)

White or Caucasian × Democrat 168.2(95.54)

White or Caucasian × Republican -14.72(104.3)

Red 144.9(82.31)

Swing 76.50(64.10)

White or Caucasian × Red -212.3(115.3)

White or Caucasian × Swing -31.55(91.68)

Constant 1763.5 1830.7 1732.9(69.71) (49.26) (50.37)

Observations 1167 1171 1193Notes: The table presents the estimates from an OLS regression of Points in the race salient altruism treatments on theemployer’s race. The omitted employer is the Black employer. Column (1), (2), and (3) test for the heterogeneity intreatment effects by income, political affiliation, and the voting pattern of the most lived state (red, blue, or swing) ofthe worker respectively. The omitted categories for income, political affiliation, and state voting pattern are less than$20,000, democrat, and blue state. Standard errors in parentheses.

47

Table C6: Heterogeneous Treatment Effects - Reciprocity Treatments

(1) (2) (3)Gender Age Education

White or Caucasian 42.95 -0.996 58.85(52.94) (112.1) (113.7)

Male 171.1(57.27)

White or Caucasian ×Male -113.1(81.06)

Age 25 - 34 -100.2(89.40)

Age 35 - 44 -119.0(95.48)

Age 45 - 54 -216.9(109.7)

Age 55 - 64 -319.6(131.3)

Age 65 or older -520.6(167.4)

White or Caucasian × Age 25 - 34 24.34(129.5)

White or Caucasian × Age 35 - 44 -4.573(136.9)

White or Caucasian × Age 45 - 54 -63.38(156.9)

White or Caucasian × Age 55 - 64 -110.6(180.6)

White or Caucasian × Age 65 or older 303.7(243.0)

Some college 190.1(93.43)

College graduate 20.81(89.36)

Graduate or professional degree -24.39(100.4)

White or Caucasian × Some college -117.5(135.9)

White or Caucasian × College graduate -63.39(129.5)

White or Caucasian × Graduate or professional degree -42.49(149.5)

Constant 1731.1 1941.5 1747.4(37.40) (77.08) (77.05)

Observations 1170 1176 1178Notes: The table presents the estimates from an OLS regression of Points in the race salient reciprocity treatments onthe employer’s race. The omitted employer is the Black employer. Column (1), (2), and (3) test for the heterogeneityin treatment effects by gender, age, and education respectively. The omitted categories for gender, age, and educationare female, age between 18 and 24, and high school or less. Standard errors in parentheses.

48

Table C7: Heterogeneous Treatment Effects - Reciprocity Treatments

(1) (2) (3)Income Political Affiliation State Voting Pattern

White or Caucasian 85.03 49.63 -67.15(100.9) (62.58) (72.69)

Income $20,000 - $44,999 15.20(88.56)

Income $45,000 - $99,999 100.2(86.06)

Income $100,000 - $149,999 98.32(116.9)

Income $150,000+ 29.43(219.8)

White or Caucasian × Income $20,000 - $44,999 37.44(123.4)

White or Caucasian × Income $45,000 - $99,999 -195.8(119.2)

White or Caucasian × Income $100,000 - $149,999 -325.6(170.2)

White or Caucasian × Income $150,000+ -63.35(284.5)

Independent 134.1(66.38)

Republican 76.13(71.70)

White or Caucasian × Independent -127.8(94.79)

White or Caucasian × Republican -35.17(101.7)

Red 12.30(82.52)

Swing 16.26(64.50)

White or Caucasian × Red 89.20(117.8)

White or Caucasian × Swing 86.19(91.82)

Constant 1752.3 1729.4 1793.2(73.27) (44.43) (50.53)

Observations 1161 1149 1179Notes: The table presents the estimates from an OLS regression of Points in the race salient reciprocity treatments onthe employer’s race. The omitted employer is the Black employer. Column (1), (2), and (3) test for the heterogeneity intreatment effects by income, political affiliation, and the voting pattern of the most lived state (red, blue, or swing) ofthe worker respectively. The omitted categories for income, political affiliation, and state voting pattern are less than$20,000, democrat, and blue state. Standard errors in parentheses.

49

Table C8: Social Preference Treatments - Robustness

Altruism Reciprocity(1) (2) (3) (4) (5) (6)

White or Caucasian -90.28 -84.51 -124.1 -5.379 -1.803 -70.48(39.92) (40.77) (99.48) (40.01) (40.89) (102.3)

Constant 1798.4 1822.5 1709.2 1803.6 1772.0 1753.7(28.12) (294.0) (302.1) (28.28) (289.2) (300.0)

Demographic Controls No Yes Yes No Yes YesEmployer Fixed Effects No No Yes No No YesObservations 1193 1138 1138 1179 1126 1126

Notes: The table presents the estimates from an OLS regression of Points in the race salient social preference treat-ments on the employer’s race. The omitted category is the Black employer. Demographic controls include age, gender,education, income, political affiliation and the voting pattern of the most lived state (red, blue, or swing) of the worker.There are total of 12 employer fixed effects for each of altruism and reciprocity treatments. Standard errors in paren-theses.

Table C9: Social Preference Treatments - Robustness, Employer Race Correctly Perceived

Altruism Reciprocity(1) (2) (3) (4) (5) (6)

White or Caucasian -92.45 -92.45 -25.12 -15.02 -5.901 -33.61(41.94) (42.97) (106.6) (42.88) (43.80) (108.6)

Constant 1807.7 1703.7 1487.5 1818.8 1788.0 1739.3(30.53) (316.5) (325.1) (31.63) (290.1) (303.7)

Demographic Controls No Yes Yes No Yes YesEmployer Fixed Effects No No Yes No No YesObservations 1051 1000 1000 1031 986 986

Notes: The table presents the estimates from an OLS regression of Points in the race salient social preference treat-ments on the employer’s race for workers who could correctly perceive the race of the employer. The omitted categoryis the Black employer. Demographic controls include age, gender, education, income, political affiliation and the vot-ing pattern of the most lived state (red, blue, or swing) of the worker. There are total of 12 employer fixed effects foreach of altruism and reciprocity treatments. Standard errors in parentheses.

50

Table C11: Balance Checks - Altruism Treatments

(1) (2) (3)Altruism - Neutral Altruism - Black Altruism - White

GenderFemale -0.01 (0.01) -0.00 (0.01) -0.00 (0.01)

Age25-30 -0.01 (0.01) -0.01 (0.01) 0.01 (0.01)31-40 0.00 (0.01) -0.01 (0.01) 0.01 (0.01)41-50 0.01 (0.02) -0.02 (0.02) 0.02 (0.02)51-64 -0.01 (0.02) -0.01 (0.02) 0.02 (0.02)65 and over 0.00 (0.03) -0.02 (0.03) -0.00 (0.03)

EducationHigh school or equivalent 0.04 (0.05) -0.02 (0.05) -0.01 (0.05)Some college 0.04 (0.05) -0.03 (0.05) 0.01 (0.05)College graduate 0.04 (0.05) -0.02 (0.05) 0.02 (0.05)Graduate or professional degree 0.05 (0.05) -0.01 (0.05) 0.00 (0.05)

Income$20,000 - $44,999 0.01 (0.01) 0.00 (0.01) -0.01 (0.01)$45,000 - $99,999 0.01 (0.01) -0.00 (0.01) -0.01 (0.01)$100,000 - $149,999 -0.00 (0.02) -0.01 (0.02) -0.01 (0.02)$150,000+ -0.00 (0.02) -0.01 (0.02) -0.02 (0.02)

Most lived US StateBlue -0.00 (0.01) -0.00 (0.01) -0.00 (0.01)Red 0.01 (0.01) -0.01 (0.01) 0.00 (0.01)

PartyDemocrat -0.00 (0.01) -0.00 (0.01) 0.00 (0.01)Republican -0.02 (0.01) -0.00 (0.01) 0.01 (0.01)

Constant 0.07 (0.05) 0.14 (0.05) 0.09 (0.05)Observations 5945 5945 5945R2 0.002 0.001 0.002F 0.68 0.37 0.69

Notes: In Columns 1-3 we regress the assignment to different conditions (Altruism Treatments) on the worker demo-graphics. Standard errors in parentheses.

51

Table C12: Balance Checks - Reciprocity Treatments

(1) (2) (3)Reciprocity - Neutral Reciprocity - Black Reciprocity - White

GenderFemale 0.01 (0.01) -0.01 (0.01) -0.01 (0.01)

Age25-30 -0.03 (0.01) -0.01 (0.01) -0.00 (0.01)31-40 -0.03 (0.01) -0.02 (0.01) 0.01 (0.01)41-50 -0.05 (0.02) -0.02 (0.02) 0.00 (0.02)51-64 -0.03 (0.02) -0.03 (0.02) 0.02 (0.02)65 and over -0.08 (0.03) 0.02 (0.03) 0.02 (0.03)

EducationHigh school or equivalent 0.04 (0.05) -0.03 (0.05) -0.01 (0.05)Some college 0.03 (0.05) -0.03 (0.05) 0.01 (0.05)College graduate 0.03 (0.05) -0.04 (0.05) 0.01 (0.05)Graduate or professional degree 0.03 (0.05) -0.02 (0.05) -0.01 (0.05)

Income$20,000 - $44,999 0.01 (0.01) 0.02 (0.01) 0.00 (0.01)$45,000 - $99,999 0.02 (0.01) 0.02 (0.01) 0.01 (0.01)$100,000 - $149,999 0.02 (0.02) 0.02 (0.02) -0.01 (0.02)$150,000+ 0.04 (0.02) -0.03 (0.02) 0.00 (0.02)

Most lived US StateBlue -0.00 (0.01) -0.00 (0.01) -0.01 (0.01)Red -0.01 (0.01) -0.01 (0.01) -0.01 (0.01)

PartyDemocrat -0.01 (0.01) 0.00 (0.01) 0.01 (0.01)Republican 0.01 (0.01) -0.01 (0.01) -0.01 (0.01)

Constant 0.09 (0.06) 0.14 (0.05) 0.10 (0.05)Observations 5945 5945 5945R2 0.004 0.002 0.003F 1.32 0.76 0.99

Notes: In Columns 1-3 we regress the assignment to different conditions (Reciprocity Treatments) on the workerdemographics. Standard errors in parentheses.

52

Table C13: Overall Productivity by Demographics

(1)Points

GenderFemale -135.42 (17.77)

Age25-30 -26.53 (29.58)31-40 -83.18 (31.39)41-50 -126.63 (35.09)51-64 -257.55 (40.42)65 and over -356.25 (58.48)

EducationSome college 1.78 (29.12)College graduate -96.92 (28.06)Graduate or professional degree -97.23 (32.92)Prefer not to answer -1260.07 (472.82)

Income$20,000 - $44,999 33.00 (25.98)$45,000 - $99,999 40.73 (26.24)$100,000 - $149,999 84.57 (37.01)$150,000+ 91.32 (54.65)

PartyDemocrat -60.48 (20.59)Republican -25.35 (22.64)

Most lived US StateBlue -47.50 (20.02)Red -13.10 (23.06)

Constant 2074.68 (38.74)Observations 5945R2 0.034F 11.68

Notes: The table presents the estimates of an OLS regression of points scored on worker demographics. Standarderrors in parentheses.

53